Self-adaptive turning control method for power station inspection robot in narrow space

Abstract

The invention relates to the technical field of deep learning, and discloses a power station inspection robot narrow space adaptive turning control method and system. The method comprises the following steps: synchronously acquiring data of a laser radar, an inertial measurement unit and a fisheye camera through multi-modal sensing; constructing an independent feature extraction channel to generate each modal feature vector; a cross-modal confidence degree dynamic evaluation mechanism is introduced, and the reliability of each sensor is quantified in real time; executing confidence gating fusion to form anti-noise state representation; and outputting a future motion sequence through a time sequence prediction network, generating PID control parameters through a decoding network, and finally smoothly resolving a steering and linear speed instruction. The sensing robustness and the turning stability of the robot in a narrow and strong interference environment are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of deep learning technology, specifically relating to an adaptive turning control method for power plant inspection robots in confined spaces. Background Technology

[0002] With the rapid development of smart grids and unmanned operation and maintenance technologies, the application of power plant inspection robots in high-risk or complex industrial scenarios such as substations and distribution rooms is becoming increasingly widespread. These environments typically feature narrow passages, dense equipment, and complex electromagnetic environments, placing extremely high demands on the robot's motion control precision and environmental adaptability. Inspection robots need to perform stable movement, precise obstacle avoidance, and agile turning within limited spaces, which relies heavily on environmental perception and pose estimation information provided by multi-sensor fusion.

[0003] However, in actual operation, the strong electromagnetic interference generated by the high-voltage equipment, switch operation and high-power cables inside the power station can easily couple to the sensor signal link, resulting in a significant increase in sudden noise in the raw data, which seriously undermines the reliability of the sensing system.

[0004] Adaptive turning control in confined spaces is a key technology for achieving efficient operation of inspection robots. This process requires the robot to dynamically adjust its turning angle and speed based on real-time perception results to complete path switching without colliding with surrounding equipment. Traditional control methods often employ rule-based logic judgments or feedback controllers with fixed parameters, which are highly sensitive to the purity of the input data. When multi-source sensors experience data distortion due to electromagnetic interference, the system struggles to effectively distinguish between real obstacles and noise artifacts, easily leading to misjudgments.

[0005] In existing technologies, some solutions attempt to introduce deep learning models for end-to-end turning decisions, but their feature extraction layers generally lack the ability to explicitly model and suppress electromagnetic noise. Under conditions of sudden noise increases, the internal feature representation of the model is easily dominated by interference, leading to violent fluctuations in the output control commands, manifested as directional jitter, trajectory deviation, or even motion instability during the turning process.

[0006] Furthermore, while conventional attention mechanisms can focus on key areas, they fail to address the specific characteristics and spatial distribution of noise, thus failing to maintain robustness in feature extraction under strong interference. These shortcomings make it difficult for existing inspection robots to achieve safe, smooth, and reliable autonomous turning in the confined and highly electromagnetically interfered environment of power plants. Therefore, an adaptive turning control method with anti-interference feature extraction and dynamic noise suppression capabilities is urgently needed. Summary of the Invention

[0007] The technical problem to be solved by this invention is to provide an adaptive turning control method for power plant inspection robots in confined spaces. This method aims to overcome the technical defects in the prior art where inspection robots in confined spaces with strong electromagnetic interference are prone to sudden strong noise during multi-sensor data fusion, which leads to incorrect judgments by the deep learning model and causes the robot's turning motion to jitter, become unstable, or even collide.

[0008] To address the aforementioned technical problems, this invention provides an adaptive turning control method for a power plant inspection robot in confined spaces, the method comprising the following steps: The multimodal perception system mounted on the power station inspection robot simultaneously acquires the robot's three-dimensional point cloud data from the lidar, angular velocity and linear acceleration data from the inertial measurement unit, and panoramic image data from the fisheye camera in the current environment. Parallel independent feature extraction channels for different data modalities are constructed to extract depth features from the lidar 3D point cloud data, the inertial measurement unit angular velocity and linear acceleration data, and the fisheye camera panoramic image data, respectively, generating corresponding spatial geometric feature vectors, motion dynamic temporal feature vectors, and visual environment texture feature vectors. A dynamic evaluation mechanism for cross-modal data confidence is established. For any data modality, the mechanism uses the feature vectors of all other data modalities as input, generates a theoretical prediction value for the target modal data through a prediction model, calculates the deviation metric between the theoretical prediction value and the actual observation value of the target modal data, and finally converts the deviation metric into a real-time confidence score characterizing the reliability of the target modal data through a nonlinear mapping function. A confidence-gated feature fusion operation is performed, in which the spatial geometric feature vector, the motion dynamic temporal feature vector, and the visual environment texture feature vector are weighted with their respective real-time confidence scores to generate a weighted feature vector. Then, all weighted feature vectors are concatenated to form a fusion state representation that can dynamically suppress the influence of noise modes. The fusion state representation is input into a temporal state evolution prediction network in a time sequence. The network models the robot's current and historical fusion states and outputs a prediction of the robot's motion state sequence for multiple consecutive time steps in the future. Based on the prediction of the future motion state sequence, a control command decoding network is used to calculate and generate steering angular velocity and linear velocity commands to drive the power station inspection robot to perform smooth and precise turning movements.

[0009] As one embodiment of the present invention, the simultaneous acquisition of lidar 3D point cloud data, inertial measurement unit angular velocity and linear acceleration data, and fisheye camera panoramic image data of the robot in the current environment through the multimodal perception system mounted on the power station inspection robot specifically includes: A 32-line lidar mounted on the top of the robot acquires three-dimensional point cloud data within a 360° field of view around the robot at a frequency of 10 Hz. The operating wavelength of the lidar is 905 nanometers. The robot's three-axis angular velocity data and three-axis acceleration data are acquired at a frequency of 200 Hz through a six-axis microelectromechanical system inertial measurement unit built into the robot's main control unit. The angular velocity range of the inertial measurement unit is ±2000° per second, and the acceleration range is ±16 times the gravitational acceleration. Panoramic image data is acquired at a frequency of 30 Hz using two fisheye cameras with a field of view of 190° symmetrically installed at the front and rear of the robot. The image sensor resolution of the camera is 1920×1080. All sensor data is timestamped and synchronized using a precise time protocol within the robot, ensuring that the data is aligned to within one millisecond in time.

[0010] Furthermore, the construction of parallel, independent feature extraction channels for different data modalities specifically includes: For the three-dimensional point cloud data of the lidar, a three-dimensional sparse convolutional network is used for processing. The network directly extracts spatial geometric feature vectors representing information such as the distribution of spatial obstacles, channel width, and wall normal vectors from the original point cloud through asymmetric convolutional kernels and submanifold sparse convolutional layers. The angular velocity and linear acceleration data of the inertial measurement unit are processed by a one-dimensional temporal convolutional network. The network contains multiple convolutional kernel groups with different dilation rates to capture the dynamic characteristics of robot motion at different time scales and is highly sensitive to spike noise caused by electromagnetic interference, thereby extracting the temporal feature vector of motion dynamics. For the panoramic image data from the fisheye camera, the image is first distorted using a preset camera intrinsic parameter model. Then, the distorted image is input into a lightweight convolutional neural network. The network extracts visual environment texture feature vectors that represent information such as ground texture, device markings, and lighting changes through depthwise separable convolution.

[0011] As one embodiment of the present invention, the establishment of a dynamic evaluation mechanism for cross-modal data confidence specifically includes: The prediction model is a multilayer perceptron network, which has three hidden layers, each with 256 neurons, and uses a linear rectified unit as the activation function. For the confidence assessment of the inertial measurement unit mode, the spatial geometric feature vector and the visual environment texture feature vector are concatenated and then input into the multilayer perceptron network. The network outputs the theoretical prediction value of the six-axis data of the inertial measurement unit at that moment. The deviation metric is the mean square error between the theoretical predicted value and the actual observed value of the inertial measurement unit; The nonlinear mapping function is a Gaussian radial basis function, which is in the form of e multiplied by negative α and the square power of the deviation measure, where α is a preset attenuation coefficient used to map the mean square error to a confidence score in the interval of 0 to 1. The larger the error, the closer the score is to 0.

[0012] For confidence assessment of LiDAR mode and fisheye camera mode, a similar prediction and bias calculation process is performed.

[0013] Furthermore, the specific steps of performing the confidence-gated feature fusion operation include: The original feature vectors extracted from each modality are multiplied element-wise with the corresponding real-time confidence scores, which range from 0 to 1, to obtain the weighted feature vectors. When sensor data of a certain modality is strongly disturbed, causing its observed value to deviate greatly from the predicted value of other modalities, its corresponding confidence score will approach 0. Thus, after the product operation, the contribution of the feature vector of that modality in the fusion process is effectively suppressed. All weighted feature vectors are concatenated along the feature dimension to form a single, dimensionally integrated fusion state representation. This representation contains effective information from multiple modalities and achieves adaptive shielding of noise sources through a confidence gating mechanism.

[0014] In one embodiment of the present invention, the temporal state evolution prediction network is a gated recurrent unit network comprising two stacked recurrent layers, each layer having 512 hidden units. The network receives the fused state representation at the current moment and updates its internal hidden states, which encode the robot's complete motion history and environmental perception history from startup to the current moment. The network output is a sequential prediction of the robot's six-DOF pose for the next ten time steps based on the current hidden state.

[0015] Furthermore, the control command decoding network is a fully connected feedforward network, whose input is the final hidden state of the pose prediction sequence for the next ten time steps output by the gated recurrent unit network. The decoding network does not directly output motor control commands, but instead outputs a five-dimensional control parameter vector, corresponding to the proportional gain, integral gain, and derivative gain of the proportional-integral-derivative (PID) controller, as well as the robot's target linear velocity and target steering angular velocity. The final robot chassis drive commands, i.e., the actual steering angular velocity and linear velocity commands applied to the drive wheels, are calculated and generated using a deterministic PID control law function, utilizing the control parameters output by the decoding network and the error between the robot's current pose and the target pose. This design separates the decision-making function of the deep learning model from the stability of the underlying motion control, using a neural network to dynamically adjust the parameters of the classical controller to ensure smooth control under complex dynamic conditions.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By creating a dynamic evaluation mechanism for cross-modal data confidence, this method can quantify the reliability of each sensor data stream in real time and online. When any sensor generates abnormal data due to electromagnetic interference, the system can immediately reduce the weight of that data source in subsequent decisions, thereby fundamentally solving the problem of noise data polluting the overall state estimation in traditional fusion methods and greatly enhancing the robot's perception robustness in strong electromagnetic interference environments.

[0017] 2. This method uses confidence gating for feature fusion, which achieves adaptive and dynamic "soft isolation" of noise sources. Compared with the hard switching scheme that detects faults based on fixed thresholds and crudely eliminates the entire sensor, this method can retain the effective information in the system to the greatest extent. Even when the data quality of some sensors deteriorates, it can maintain the continuity and integrity of the overall perception.

[0018] 3. By combining a temporal state evolution prediction network with a decoding network for output proportional-integral-derivative (PI-DE) control parameters, this method decouples high-level temporal decision-making from low-level motion execution. The neural network is responsible for making macroscopic and predictive decisions based on complex perceptual information, generating optimal control strategy parameters. The final control commands are generated by a structurally stable and predictable PI-DE controller, effectively avoiding output abrupt changes and jitter that may occur in end-to-end models. This ensures the continuity and smoothness of the robot's turning movements in confined spaces, significantly improving operational safety and reliability.

[0019] 4. This invention constructs a complete end-to-end processing link from multimodal raw data to smooth control instructions. All modules are designed as differentiable structures, enabling the entire system to be trained through global optimization. This achieves deep collaboration between perception, decision-making, and control, ensuring optimal overall performance in complex, dynamic, and confined space scenarios. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall technical solution architecture of the adaptive turning control method for power plant inspection robots in confined spaces proposed in this invention; Figure 2 This is a schematic diagram illustrating the core principle framework of the cross-modal data confidence dynamic evaluation mechanism in this invention; Figure 3 This is a flowchart illustrating the logical process of synchronous acquisition of multimodal sensing data and independent feature extraction in this invention. Figure 4 This is a logical flowchart of the confidence-gated feature fusion operation in this invention; Figure 5 This is a logical flowchart of the timing state evolution prediction and control instruction decoding in this invention; Figure 6 This is a schematic diagram of the multi-level interaction relationship and data flow between the power station inspection robot, the multimodal perception system, and the control execution unit in this invention. Detailed Implementation

[0021] Please refer to Figures 1 to 6 This invention provides an adaptive turning control method for power plant inspection robots in confined spaces. This method constructs an end-to-end adaptive control link by building a multimodal perception system, independent feature extraction channels, a cross-modal confidence dynamic evaluation mechanism, confidence-gated feature fusion operations, a temporal state evolution prediction network, and a control command decoding network. This ensures that the robot can achieve smooth, stable, and safe turning operations in confined spaces.

[0022] The method includes the following steps: S1, through the multimodal perception system mounted on the power station inspection robot, simultaneously acquire the robot's three-dimensional point cloud data from the lidar, the angular velocity and linear acceleration data from the inertial measurement unit, and the panoramic image data from the fisheye camera in the current environment. S2, construct parallel independent feature extraction channels for different data modalities, and perform depth feature extraction on the lidar 3D point cloud data, the inertial measurement unit angular velocity and linear acceleration data, and the fisheye camera panoramic image data respectively, generating their corresponding spatial geometric feature vectors, motion dynamic temporal feature vectors, and visual environment texture feature vectors; S3. Establish a dynamic evaluation mechanism for cross-modal data confidence. For any data modality, the mechanism uses the feature vectors of all other data modalities as input, generates a theoretical prediction value for the target modal data through a prediction model, calculates the deviation measure between the theoretical prediction value and the actual observation value of the target modal data, and finally converts the deviation measure into a real-time confidence score characterizing the reliability of the target modal data through a nonlinear mapping function. S4, perform confidence-gated feature fusion operation, weight the spatial geometric feature vector, the motion dynamic temporal feature vector and the visual environment texture feature vector with their respective real-time confidence scores to generate weighted feature vectors, and then concatenate all weighted feature vectors to form a fusion state representation that can dynamically suppress the influence of noise modes; S5, the fusion state representation is input into a temporal state evolution prediction network according to the time sequence. The network models the current and historical fusion states of the robot and outputs a prediction of the robot's motion state sequence for multiple consecutive time steps in the future. S6. Based on the prediction of the future motion state sequence, a control command decoding network is used to calculate and generate steering angular velocity and linear velocity commands for driving the power station inspection robot to perform smooth and precise turning movements.

[0023] In step S1, the multimodal perception system mounted on the power station inspection robot simultaneously acquires the robot's three-dimensional point cloud data from the lidar, angular velocity and linear acceleration data from the inertial measurement unit, and panoramic image data from the fisheye camera in the current environment.

[0024] Specifically, a 32-line lidar mounted on the top of the robot acquires three-dimensional point cloud data within a 360° field of view around the robot at a frequency of 10 Hz. The operating wavelength of the lidar is 905 nanometers. The robot's three-axis angular velocity data and three-axis acceleration data are acquired at a frequency of 200 Hz through a six-axis microelectromechanical system inertial measurement unit built into the robot's main control unit. The angular velocity range of the inertial measurement unit is ±2000° per second, and the acceleration range is ±16 times the gravitational acceleration. Panoramic image data is acquired at a frequency of 30 Hz using two fisheye cameras with a 190° field of view symmetrically mounted at the front and rear of the robot. The image sensor resolution of the camera is 1920×1080. All sensor data is timestamped and synchronized through a precise time protocol inside the robot to ensure that the data alignment accuracy in the time dimension reaches one millisecond.

[0025] The synchronization mechanism combines hardware triggering with clock calibration. The main control unit sends a unified sampling trigger signal. After receiving the signal, each sensor module immediately latches the current data and adds a local high-precision timestamp. Then, the data with different sampling rates are uniformly interpolated to the preset reference time axis through the timestamp alignment algorithm to form a strictly aligned multimodal data frame.

[0026] In step S2, parallel independent feature extraction channels for different data modalities are constructed to extract depth features from the lidar 3D point cloud data, the inertial measurement unit angular velocity and linear acceleration data, and the fisheye camera panoramic image data, generating corresponding spatial geometric feature vectors, motion dynamic temporal feature vectors, and visual environment texture feature vectors.

[0027] For the aforementioned 3D point cloud data from the LiDAR, a 3D sparse convolutional network is used for processing. The network directly extracts spatial geometric feature vectors representing information such as the distribution of spatial obstacles, channel width, and wall normal vectors from the original point cloud through asymmetric convolutional kernels and submanifold sparse convolutional layers.

[0028] The network first converts the point cloud into sparse tensors, with each voxel recording whether a point exists and its average coordinate offset. Then, it passes through four sub-manifold sparse convolutional layers in sequence, with each convolutional kernel having a size of 3×3×3 and the number of channels being 32, 64, 128, and 256, respectively. The activation function is a leaky linear rectified unit. Finally, a 512-dimensional spatial geometric feature vector is generated through a global max pooling operation.

[0029] The angular velocity and linear acceleration data of the inertial measurement unit are processed using a one-dimensional temporal convolutional network. The network contains three sets of convolutional kernels with different dilation rates to capture the dynamic characteristics of robot motion at different time scales and has high sensitivity to spike noise caused by electromagnetic interference, thereby extracting the temporal feature vector of motion dynamics.

[0030] The network receives a time-series window of length 200, with each frame containing six-dimensional data. This data is sequentially passed through one-dimensional convolutional layers with dilation rates of 1, 2, and 4. Each layer outputs 128 channels, with a kernel size of 3 and a linear rectified unit activation function. Finally, global average pooling is used to generate a 256-dimensional motion dynamic temporal feature vector. This is specifically designed for the fisheye camera panoramic image data. First, the image is distorted using a pre-defined camera intrinsic parameter model, which includes focal length, principal point coordinates, and radial distortion coefficients. The distorted image is then cropped into two symmetrical regions to eliminate ground reflection interference. Subsequently, the distorted image is input into a lightweight convolutional neural network, which extracts visual environment texture feature vectors representing ground texture, device markings, and lighting changes through depthwise separable convolution.

[0031] The network contains five depthwise separable convolutional blocks, each consisting of depthwise convolution and pointwise convolution, with channel numbers of 32, 64, 128, 256, and 512 respectively. Finally, a 512-dimensional visual environment texture feature vector is generated through global average pooling.

[0032] In step S3, a dynamic evaluation mechanism for cross-modal data confidence is established. For any data modality, the mechanism uses the feature vectors of all other data modalities as input, generates a theoretical prediction value for the target modal data through a prediction model, calculates the deviation metric between the theoretical prediction value and the actual observation value of the target modal data, and finally converts the deviation metric into a real-time confidence score characterizing the reliability of the target modal data through a nonlinear mapping function.

[0033] The prediction model is a multilayer perceptron network with three hidden layers, each containing 256 neurons, and using a linear rectified unit as the activation function. For the confidence assessment of the inertial measurement unit (IMU) mode, the spatial geometric feature vector and the visual environment texture feature vector are concatenated and input into the IMU network. The network outputs the theoretical prediction value corresponding to the IMU's six-axis data at that moment. The deviation metric is the mean square error between the theoretical predicted value and the actual observed value of the inertial measurement unit; the nonlinear mapping function is a Gaussian radial basis function, and its form is: α is a preset attenuation coefficient used to map the mean square error to a confidence score in the range of 0 to 1. The larger the error, the closer the score is to 0. Similar prediction and bias calculation processes are performed for the confidence assessment of LiDAR mode and fisheye camera mode.

[0034] Specifically, for the LiDAR mode, the motion dynamic temporal feature vector and the visual environment texture feature vector are concatenated and then input into another multilayer perceptron network to predict the statistical summary features of the LiDAR point cloud, such as the distance to the nearest obstacle and the direction of the channel centerline. The cosine distance between the predicted value and the actual extracted spatial geometric feature vector is calculated as a deviation metric. For the fisheye camera modality, the spatial geometric feature vector and the motion dynamic temporal feature vector are concatenated and input into a third multilayer perceptron network to predict key semantic regions in the image, such as device label locations and ground arrow directions. The cross-entropy loss between the predicted semantic map and the actual visual features is calculated as a bias metric. All confidence scores are normalized using the same Gaussian radial basis function to ensure consistent numerical ranges.

[0035] In step S4, a confidence-gated feature fusion operation is performed, in which the spatial geometric feature vector, the motion dynamic temporal feature vector, and the visual environment texture feature vector are weighted with their respective real-time confidence scores to generate a weighted feature vector. Then, all weighted feature vectors are concatenated to form a fusion state representation that can dynamically suppress the influence of noise modes.

[0036] Specifically, the original feature vectors extracted from each modality are multiplied element-wise with the corresponding real-time confidence scores, which range from 0 to 1, to obtain a weighted feature vector. When the sensor data of a certain modality is strongly disturbed, causing its observed value to deviate significantly from the predicted value of other modalities, its corresponding confidence score will approach 0. Thus, after the multiplication operation, the contribution of the feature vector of that modality to the fusion process is effectively suppressed. All weighted feature vectors are concatenated along their feature dimensions to form a single, dimensionally integrated fused state representation. This representation not only contains effective information from multiple modalities but also adaptively masks noise sources through a confidence-gated mechanism. The fused state representation has 1280 dimensions, with spatial geometric feature vectors contributing 512 dimensions, motion dynamic temporal feature vectors contributing 256 dimensions, and visual environment texture feature vectors contributing 512 dimensions. This fused representation serves as the input unit for subsequent temporal modeling, transmitting only the confidence-filtered effective information at each time step.

[0037] In step S5, the fusion state representation is input into a temporal state evolution prediction network in a time sequence. The network models the current and historical fusion states of the robot and outputs a prediction of the robot's motion state sequence for multiple consecutive time steps in the future.

[0038] The temporal state evolution prediction network is a gated recurrent unit network containing two stacked recurrent layers, each with 512 hidden units. The network receives the fused state representation at the current moment and updates its internal hidden states, which encode the robot's complete motion history and environmental perception history from startup to the current moment.

[0039] The network outputs a sequential prediction of the robot's six-DOF pose for the next ten time steps based on the current hidden state. Each time step's pose prediction includes three-dimensional translation coordinates and three-dimensional Euler angles, for a total of six dimensions.

[0040] During the training phase, the network is optimized using a teacher-mandated strategy, which uses real historical states as input to predict the state at the next moment and calculates the mean squared error loss against the real labels. During the inference phase, the network employs an autoregressive approach, using the current prediction as input for the next moment to progressively generate the future trajectory. This design enables the robot to proactively plan turning paths, avoiding collision risks caused by localized perception blind spots.

[0041] In step S6, based on the predicted future motion state sequence, a control command decoding network is used to calculate and generate steering angular velocity and linear velocity commands for driving the power station inspection robot to perform smooth and precise turning movements. The control command decoding network is a fully connected feedforward network, whose input is the final hidden state of the pose prediction sequence for the next ten time steps output by the gated recurrent unit network. The decoding network does not directly output motor control commands, but instead outputs a five-dimensional control parameter vector, corresponding to the proportional gain, integral gain, and derivative gain of the proportional-integral-derivative controller, as well as the robot's target linear velocity and target steering angular velocity.

[0042] The final robot chassis drive commands, namely the steering angular velocity and linear velocity commands actually applied to the drive wheels, are calculated and generated using a deterministic proportional-integral-derivative control law function, utilizing the control parameters output by the decoding network and the error between the robot's current pose and the target pose.

[0043] This design separates the decision-making function of the deep learning model from the stability of the underlying motion control, using a neural network to dynamically adjust the parameters of the classical controller to ensure smooth control under complex dynamic conditions. The proportional-integral-derivative control law function is defined as follows: ; in, To control the output, The error between the current pose and the target pose. , , These are the proportional, integral, and derivative gains output from the decoding network, respectively. The target linear velocity and target steering angular velocity are used as reference trajectory inputs to the controller, and the error term... This includes lateral deviation, heading angle deviation, and their derivatives. This control architecture ensures that even with minor fluctuations in the neural network output, the underlying controller maintains the continuity and physical plausibility of output commands, effectively suppressing jitter.

[0044] The above method steps S1 to S6 constitute a closed-loop control process, executing a complete cycle every 20 milliseconds. In the event of sudden electromagnetic interference, for example, if the inertial measurement unit outputs an abnormal angular velocity value due to the action of a nearby high-voltage switch, the cross-modal confidence assessment mechanism will detect the significant deviation between the modal data and the lidar and visual prediction results, reducing its confidence score to near 0, thereby automatically shielding the noise source during the feature fusion stage.

[0045] At this point, the system primarily relies on spatial geometry and visual texture features for state estimation and control decisions, maintaining the stability of the turning maneuver. Once the disturbance disappears, the confidence score automatically recovers, and the system reintroduces inertial data to improve dynamic response accuracy. The entire process requires no manual intervention and does not rely on preset fault thresholds, achieving adaptive and soft isolation from noise.

[0046] The method described in this embodiment fundamentally solves the reliability problem of sensor data fusion under strong electromagnetic interference through the collaborative design of multimodal sensing, dynamic confidence assessment, gating fusion, timing prediction and parametric decoding. By decoupling high-level decision-making from low-level control, it ensures the smoothness and security of output commands and is suitable for inspection spaces with dense metal structures and high-frequency electromagnetic radiation, such as power plants and substations.

Claims

1. A power plant inspection robot narrow space adaptive turning control method, characterized in that, The method comprises the following steps: Synchronously acquiring laser radar three-dimensional point cloud data, inertial measurement unit angular velocity and linear acceleration data, and fisheye camera panoramic image data of the robot in the current environment through a multi-modal perception system mounted on the power station inspection robot; Constructing parallel independent feature extraction channels for different data modalities, respectively performing deep feature extraction on the laser radar three-dimensional point cloud data, the inertial measurement unit angular velocity and linear acceleration data, and the fisheye camera panoramic image data to generate corresponding spatial geometric feature vectors, motion dynamic time sequence feature vectors and visual environment texture feature vectors; Establishing a cross-modal data confidence dynamic evaluation mechanism, for any data modality, using the feature vectors of all other data modalities as input, generating a theoretical prediction value of the target modality data through a prediction model, calculating the deviation measure between the theoretical prediction value and the actual observation value of the target modality data, and finally converting the deviation measure into a real-time confidence score representing the reliability of the target modality data through a nonlinear mapping function; Performing confidence gating feature fusion operation, weighting the spatial geometric feature vectors, the motion dynamic time sequence feature vectors and the visual environment texture feature vectors respectively with the corresponding real-time confidence scores to generate weighted feature vectors, and then splicing all the weighted feature vectors to form a fusion state representation that can dynamically suppress the influence of noise modalities; Inputting the fusion state representation into a time sequence state evolution prediction network in time sequence, the network models the current and historical fusion state of the robot and outputs the prediction of the motion state sequence of the robot at future continuous time steps; According to the prediction of the future motion state sequence, through a control instruction decoding network, the steering angular velocity and linear velocity instructions for driving the power station inspection robot to perform smooth and accurate turning actions are calculated and generated.

2. The power station inspection robot narrow space adaptive turning control method according to claim 1, characterized in that, Synchronously acquiring laser radar three-dimensional point cloud data, inertial measurement unit angular velocity and linear acceleration data, and fisheye camera panoramic image data of the robot in the current environment through a multi-modal perception system mounted on the power station inspection robot, comprising: Acquiring three-dimensional point cloud data within a 360° field of view around the robot at a frequency of 10 Hz through a 32-line laser radar installed on the top of the robot, the working wavelength of the laser radar is 905 nanometers; Acquiring three-axis angular velocity data and three-axis linear acceleration data of the robot at a frequency of 200 Hz through a six-axis micro-electro-mechanical system inertial measurement unit built-in the robot main control unit, the angular velocity range of the inertial measurement unit is ±2000° per second, and the acceleration range is ±16 times of gravity acceleration; Acquiring panoramic image data at a frequency of 30 Hz through two fisheye cameras with a 190° field of view angle symmetrically installed on the front and rear of the robot, the image sensor resolution of the camera is 1920×1080; All sensor data are time-stamped synchronized through the precise time protocol inside the robot, ensuring that the alignment accuracy of the data in the time dimension reaches one millisecond.

3. The method of claim 2, wherein, Constructing parallel, independent feature extraction channels for different data modalities, including: For the laser radar three-dimensional point cloud data, a three-dimensional sparse convolution network is used for processing, which extracts spatial geometric feature vectors representing spatial obstacle distribution, channel width, wall normal vector, etc. through asymmetric convolution kernel and sub-manifold sparse convolution layer directly on the original point cloud; For the inertial measurement unit angular velocity and linear acceleration data, a one-dimensional time series convolution network is used for processing, which contains multiple convolution kernel groups with different expansion rates to capture the dynamic characteristics of robot motion at different time scales and has high sensitivity to spike noise caused by electromagnetic interference, thereby extracting motion dynamic time series feature vectors; For the fisheye camera panoramic image data, first, a preset camera intrinsic parameter model is used for distortion correction, and then the corrected image is input into a lightweight convolutional neural network, which extracts visual environment texture feature vectors representing ground texture, device identification, illumination changes, etc. through depth separable convolution.

4. The power station inspection robot narrow space adaptive turning control method according to claim 3, characterized in that, Establish a cross-modal data confidence dynamic evaluation mechanism, including: The prediction model is a multi-layer perceptron network, which contains three hidden layers, each with 256 neurons, and uses a linear rectifier unit as the activation function; For the confidence evaluation of the inertial measurement unit modality, the spatial geometric feature vector and the visual environment texture feature vector are concatenated and input into the multi-layer perceptron network, and the network outputs the theoretical prediction value of the inertial measurement unit six-axis data at that time; The bias measure is the mean square error between the theoretical prediction value and the actual observation value of the inertial measurement unit; The nonlinear mapping function is a Gaussian radial basis function, which is e raised to the power of -α times the square of the bias measure, where α is a preset decay coefficient used to map the mean square error to a confidence score in the range of 0 to 1. The larger the error, the closer the score to 0; For the confidence evaluation of the laser radar modality and the fisheye camera modality, similar prediction and bias calculation processes are performed.

5. The power station inspection robot narrow space adaptive turning control method according to claim 4, characterized in that, Perform confidence-gated feature fusion operations, including: Perform element-wise multiplication operations on the original feature vectors extracted from each modality and the corresponding real-time confidence scores with a value range of 0 to 1 to obtain weighted feature vectors; When the sensor data of a certain modality is strongly disturbed, causing a large deviation between its observation value and the prediction value of other modalities, its corresponding confidence score will tend to 0, so that the contribution of the feature vector of that modality in the fusion process is effectively suppressed after the multiplication operation; Concatenate all weighted feature vectors in the feature dimension to form a single, dimensionally integrated fusion state representation that contains both multi-modal effective information and adaptive noise source shielding through the confidence gating mechanism.

6. The power station inspection robot narrow space adaptive turning control method according to claim 5, characterized in that, Input the fusion state representation into a time series state evolution prediction network in time series, which models the current and historical fusion states of the robot and outputs a prediction of the robot's future motion state sequence at multiple time steps, including: The time series state evolution prediction network is a gated recurrent unit network containing two stacked recurrent layers, each with 512 hidden units; The network receives the fusion state representation at the current time step and updates its internal hidden state, whose hidden state encodes the complete motion history and environment perception history of the robot from the start to the current time step; The output of the network is a sequential prediction of the robot's 6-DOF pose at the next ten time steps based on the current hidden state.

7. The power station inspection robot narrow space adaptive turning control method according to claim 6, characterized in that, According to the prediction of the future motion state sequence, through a control instruction decoding network, the steering angular velocity and linear velocity instructions for driving the power station inspection robot to perform smooth and accurate turning actions are calculated and generated, including: The control instruction decoding network is a fully connected feedforward network, whose input is the final hidden state of the pose prediction sequence of the next ten time steps output by the gated recurrent unit network; The decoding network outputs a five-dimensional control parameter vector, corresponding to the proportional gain, integral gain, and derivative gain of the proportional-integral-derivative controller, as well as the target linear velocity and target steering angular velocity of the robot; The final robot chassis driving instructions are calculated and generated by a deterministic proportional-integral-derivative control law function using the control parameters output by the decoding network and the error between the current pose and the target pose of the robot.

8. The power station inspection robot narrow space adaptive turning control method according to claim 7, characterized in that, The three-dimensional sparse convolution network voxelizes the point cloud into a sparse tensor, with each voxel recording whether there is a point and its average coordinate offset, then sequentially passes through four sub-manifold sparse convolution layers, each with a 3x3x3 convolution kernel size and channel numbers of 32, 64, 128, and 256, respectively, with a leaky linear rectifier activation function, and finally generates a 512-dimensional spatial geometric feature vector through global max pooling.

9. The power station inspection robot narrow space adaptive turning control method according to claim 8, characterized in that, The one-dimensional time series convolution network receives a time series window of length 200, each frame containing six-dimensional data, sequentially passes through one-dimensional convolution layers with dilation rates of 1, 2, and 4, each with 128 output channels and a 3x3 convolution kernel size, with a linear rectifier activation function, and finally generates a 256-dimensional motion dynamic time series feature vector through global average pooling.

10. The power station inspection robot narrow space adaptive turning control method according to claim 9, characterized in that, The lightweight convolutional neural network contains 5 depth separable convolution blocks, each composed of a depth convolution and a pointwise convolution, with channel numbers of 32, 64, 128, 256, and 512, respectively, and finally generates a 512-dimensional visual environment texture feature vector through global average pooling.

Images